ComfyUI-FluxHallucinate

ComfyUI-FluxHallucinate

Intentionally degrade Flux dev.1's prompt adherence to recreate the loose, random, hallucinatory feel of SD1.5 / SDXL. Modern Flux is too obedient — this pack brings back generations that drift, invent, and surprise you between runs.

Why Flux feels "too coherent"

• Guidance distillation — Flux dev bakes strong prompt adherence into the guidance scalar.
• T5-XXL sequence embeddings — the real source of literal prompt-following. SDXL never had T5 (CLIP only), which is a big part of why it drifts and invents.
• Rectified-flow sampling — a near-straight path to a clean image, with little of the per-step noise wandering that ancestral SD1.5 samplers had.

These nodes attack all three.

Install

Clone into your ComfyUI custom_nodes/ and restart the server (a node-list refresh won't pick up a brand-new package):

cd ComfyUI/custom_nodes
git clone https://github.com/isam2024/ComfyUI-FluxHallucinate

No extra dependencies (pure torch).

Nodes

Flux Hallucinate — conditioning/chaos

The main creative knob. Wire it between CLIP Text Encode and the sampler/guider on the positive conditioning. It weakens / scrambles / noises the conditioning so the model interprets the prompt instead of obeying it.

| param | effect | suggested start | |---|---|---| | t5_scale | weakens literal adherence (0 ≈ ignores prompt body) | 0.65 | | embed_noise | gaussian noise on T5 embeddings → hallucinated associations | 0.20 | | token_dropout | zeroes whole tokens → dropped concepts, model fills gaps | 0.0–0.15 | | token_shuffle | permutes a fraction of tokens → scrambled word-order meaning | 0.0–0.2 | | pooled_scale | scales the global CLIP "mood" vector (SDXL leaned on this) | 1.0 | | pooled_noise | gaussian noise on the pooled CLIP vector | 0.0 | | seed | makes the degradation reproducible | — |

Flux Noise Injection — model_patches/chaos

Wire on the MODEL line before the sampler. Re-injects fresh noise every sampling step (scaled by the current sigma, so it self-limits as the image resolves). This is what recreates the ancestral-sampler wandering. Works with any sampler. Start amount ≈ 0.12.

Flux Chaos Guidance — conditioning/chaos

Sets a low Flux guidance (try 1.0–2.5) with optional per-run jitter. Low guidance relaxes the distilled adherence → dreamier, looser output. Don't stack with a separate FluxGuidance node — last one wins.

Recommended recipe (max SD1.5 chaos)

CLIP Text Encode
  → Flux Hallucinate (t5_scale 0.6, embed_noise 0.25)
  → Flux Chaos Guidance (1.5, jitter 0.5)        [conditioning]
Flux Noise Injection (0.12)                        [model]
+ ancestral/SDE sampler (euler_ancestral / dpmpp_2s_ancestral / dpmpp_sde)
+ fewer steps (~12–18)

Each layer compounds — introduce them one at a time to dial in the amount of chaos you want.

License

MIT